Ben Okonkwo: Marcus, hey — I have a question I couldn't answer over the weekend and it's been bothering me since.
Marcus Vale: Good or bad bothering?
Ben Okonkwo: Interesting bothering. What were we actually following for fifty years? Because I went back and read Gordon Moore's 1965 paper — the original — and the observation isn't about maximum transistor density. It's about the density at which cost per transistor is lowest. That's the load-bearing claim. And it's economic, not physical.
Marcus Vale: Hold on — so Moore's Law, from the beginning, was a cost-optimization insight. Not a capability ceiling.
Ben Okonkwo: From the beginning. And he was at Fairchild Semiconductor when he first deployed it — not Intel. He used it as a sales and competitive framing before it ever became a roadmap for the industry. So right from the start, it's strategic.
Marcus Vale: That's — yeah, that actually changes the framing completely. Because the version of Moore's Law that Intel later promoted through roadmaps, that the ITRS institutionalized as an industry coordination mechanism — that already feels like a second-generation artifact. A copy of something that had a very different original purpose.
Ben Okonkwo: Right — and then Moore himself revised the cadence in 1975. Changed it from annual doubling to roughly every two years. So the number in the law changed. Which is, hm, not what laws usually do.
Marcus Vale: Laws of physics don't have amendment rounds.
Ben Okonkwo: So the discomfort I'm sitting with — and I think this is what today is really about — is that if Moore's Law was always a chosen target, always adaptable, then the question isn't whether it's physically true. The question is who gets to decide when the target changes, and whether that decision has already been made without anyone announcing it.
Marcus Vale: Basically: was it ever a law, or was it always just the industry's shared calendar?
Ben Okonkwo: And if it's the latter — who tears the page off when the year ends?
Marcus Vale: The shared calendar framing — that's actually the thing I want to pull on, because it undersells what the ITRS actually did. It wasn't just a calendar. The International Technology Roadmap for Semiconductors was a planning contract across the entire supply chain. Chip makers, equipment suppliers, materials companies — everyone built their capex around the same page.
Ben Okonkwo: Right — and that's the mechanism I think most people miss. The prediction didn't come true because physics required it.
Marcus Vale: It came true because everyone bet on it coming true simultaneously.
Ben Okonkwo: Exactly — and here's the plain version of that. Imagine a group of contractors all agreeing that every house they build next year will use half the lumber at half the price. Because they all agreed, they all jointly invested in the sawmill. The price actually did fall. The prediction caused the outcome. That's what the ITRS was doing for transistors.
Marcus Vale: Self-fulfilling prophecy, but like — industrially enforced.
Ben Okonkwo: And Intel's role in that is, hm, more aggressive than it usually gets credit for. Intel wasn't just following the roadmap — they were publishing their own roadmaps and sharing them upstream with suppliers. Which meant Intel was essentially setting the cadence that every competitor then had to match. If you fell off that two-year rhythm, your customers had already planned around the next node. Defecting was commercially suicidal.
Marcus Vale: So Intel's roadmap discipline was the moat. Not the transistors — the expectation management.
Ben Okonkwo: Now here's a concrete case. A firmware engineer in Austin, 2003, speccing out a consumer router. She doesn't call Intel, she doesn't call TSMC. She opens the ITRS roadmap and reads off the transistor density and cost she can plan around — two years forward. Her entire product timeline runs on a document that exists because the industry collectively agreed to make it true.
Marcus Vale: That's the part that breaks my brain a little. The certainty wasn't coming from physics. It was coming from coordination.
Ben Okonkwo: And Intel dominated through the mid-2010s on exactly that. Relentless execution on the roadmap, ecosystem alignment — the coordination was the competitive weapon, not just the chip. But when the collective will hits a wall that the will alone can't move — that's actually the load-bearing question — what happens then?
Marcus Vale: Meaning: physics as the first thing the prophecy can't self-fulfill its way past.
Ben Okonkwo: The mechanism that made the prediction true was collective will. Not physics. So when quantum tunneling starts dominating below roughly seven nanometers — and we are already there at leading nodes — no amount of industry agreement rebuilds that cost curve. The contractors can all agree on the lumber price, but they can't agree the forest back into existence.
Marcus Vale: But the forest metaphor actually understates it — because the thing that's cracking isn't just the physics ceiling. It's that the two halves of Moore's original formulation have come apart. Density is still climbing. TSMC 2nm, Intel 18A, Samsung 2nm — those are real nodes, real transistor counts going up. But cost per transistor? That curve inverted.
Ben Okonkwo: Inverted how, specifically?
Marcus Vale: Below five nanometers — roughly 2015, 2016 — you start spending exponentially more capital to get linearly smaller gains. Transistor density per die goes up, cost per transistor goes up too. That's not slowing Moore's Law. That's reversing the part that actually mattered commercially.
Ben Okonkwo: And Rock's Law is the mechanism underneath that. Moore's Second Law — fabrication plant costs double with every new process generation. Which means the capital side is on its own exponential, going the wrong direction.
Marcus Vale: Right — and ASML is almost like the physical proof of that. Their EUV machines, extreme ultraviolet lithography, thirteen-point-five nanometer wavelength light — one machine costs hundreds of millions of dollars. That's a multi-year engineering program just to print the patterns. Prior scaling cycles didn't require anything like that.
Ben Okonkwo: Oh, that's — okay, that framing is interesting. ASML isn't a solution to the problem. It's a symptom of how expensive the problem got.
Marcus Vale: Exactly. And Intel 18A is doing RibbonFET — gate-all-around transistors — plus PowerVia, which is backside power delivery. Those are real architectural innovations. But here's the deal: you need all of that complexity now just to stay on the density curve. That's not continuity. That's — I mean, the engineering required is categorically different.
Ben Okonkwo: So the law's density component is technically alive, but only because we're substituting engineering complexity for the scaling mechanism that used to do the work automatically.
Marcus Vale: Which raises the actual question — if the original formulation was density AND falling cost, and only density is still moving... is the law alive or are we just keeping one of its two vital signs going?
Ben Okonkwo: Right, and — wait, no, I want to flag the assumption sitting under that. Moore's 1965 formulation wasn't about maximum achievable density. It was about the density at which cost per transistor hits its minimum. If cost per transistor is rising at leading-edge nodes, we're not just slowing the law. We're operating off the wrong point on Moore's own curve entirely.
Marcus Vale: We're past the optimum and nobody filed the paperwork.
Ben Okonkwo: And we've been there since roughly 2015, 2016. Nearly a decade of inverted economics running under the same branding. The part that makes this stranger — and we'll get into this — is how the industry responded to that inversion. Chiplets, 3D stacking, domain-specific chips. Which sounds like continuation but is actually a different mechanism entirely, and that distinction matters a lot.
Marcus Vale: The rebranding question. Whether those fixes are Moore's Law continuing or Moore's Law being quietly replaced by something with the same name on the door.
Ben Okonkwo: And the thing that makes the rebranding slippery — it's not dishonest, exactly. Because quantum tunneling below roughly seven nanometers is already happening. We're not predicting it. Leading nodes right now, direct tunneling is elevating leakage currents, quantum confinement is wrecking threshold voltage control. It's not a future wall. We drove through it.
Marcus Vale: Wait — we're already inside that regime?
Ben Okonkwo: Already there. Intel 18A, TSMC 2nm — those nodes exist below five nanometers. The physics isn't coming. The physics arrived. Engineers are working around it with gate-all-around transistors, with RibbonFET — wrapping the gate on all four sides of the channel so you maintain electrostatic control where a FinFET design simply breaks down.
Marcus Vale: So RibbonFET isn't an improvement on FinFET. It's a workaround for the thing FinFET couldn't survive.
Ben Okonkwo: That's — yeah, that's the cleaner framing. And Intel 18A ships with PowerVia on top of that, backside power delivery, late 2025 for high-volume manufacturing. Real engineering. But the complexity required just to maintain density... that's the tell. You need two architectural overhauls simultaneously to stay on the curve that used to happen almost automatically.
Marcus Vale: Chiplets are the same logic, right? Disaggregate the monolithic die, mix process nodes per function, high-bandwidth interconnects stitching it together — that's not scaling. That's routing around the scaling problem.
Ben Okonkwo: Right — and 3D integration stacks dies vertically, so you get density gains without shrinking the transistors in the planar sense at all. Both moves sidestep the physical limit rather than solve it. Which is genuinely impressive engineering. But Moore's original curve was commoditizable. Predictable. A firmware engineer could plan two years out without knowing anything about the fab.
Marcus Vale: Chiplets are not commoditizable. Custom design per application, custom interconnect spec. That's not a tailwind. That's a bespoke solution.
Ben Okonkwo: Which is the mechanism breaking. The original law gave you general-purpose cost improvement — economy-wide. Domain-specific processors, AI training silicon — they sustain performance-per-dollar for a narrow workload. But that's not the same contract.
Marcus Vale: And AI demand is masking that break. Because customers will pay premium for AI-optimized silicon right now regardless of whether cost-per-transistor is falling. So the investment keeps flowing, the press releases keep citing Moore's Law — but the mechanism underneath is different.
Ben Okonkwo: That's the part I find — hm. It's almost too clean. AI demand arrives exactly when the general cost curve breaks, and it props up semiconductor investment so the break stays invisible to most observers.
Marcus Vale: Convenient timing for the incumbents.
Ben Okonkwo: Now — the proponents of 'Moore's Law continues' aren't wrong that performance per dollar is still improving for specific workloads. TSMC 2nm ships, GAA transistors work, EUV from ASML keeps printing finer features. Progress is real. But I'd want to separate what's still possible from what's still predictable and general-purpose. Those are different claims.
Marcus Vale: Basically the question is whether Gordon Moore's actual formulation — cost per transistor falling on a predictable cadence — is the thing continuing. Or whether we've retired that and kept the name.
Ben Okonkwo: And I don't think either of us can answer that cleanly. The proponents point to Intel 18A and TSMC 2nm and say the cadence holds. The skeptics point to the inverted cost curve and say the name held but the mechanism didn't. Both positions have actual evidence. What I'd want to know is — what signal would make the proponents update? If cost per transistor is still rising at sub-three-nanometer nodes five years from now, does the 'continues in spirit' framing survive that?
Marcus Vale: That's the experiment. And the industry is running it right now whether it wants to or not.
Ben Okonkwo: And I don't have a clean answer for this — but somewhere right now there are product teams running five-year roadmaps, and the compute cost assumptions baked into those roadmaps are the doubling cadence. Two years. Halving cost. And nobody in the room is calling that an assumption. It's just the background. It became invisible because it was reliable for fifty years.
Marcus Vale: And the cadence has actually slipped to three to five years at leading nodes. So those roadmaps are already wrong. The teams just haven't been punished for it yet — because AI demand is covering the gap in the short term.
Ben Okonkwo: Not even lying. Just... the expectation was baked in so deep it stopped being visible as an expectation.
Marcus Vale: Which brings me back to the thing I actually can't resolve. If Moore's Law was always a choice — coordinated will, shared roadmap, industry alignment — then the question isn't whether the physics gave out. The question is whether the incentives to keep choosing it are still intact. And I'm not sure they are. TSMC, Intel, Samsung — they're all still investing in leading-edge nodes. But the economics of who can afford that table are shrinking fast. Moore's Second Law is doing its own work in the background.
Ben Okonkwo: That's — yeah. That's where I'm uneasy. Gordon Moore published a number in 1965 and the industry decided to make it true. That's an extraordinary thing. But the mechanism that made the choice credible was falling cost. If cost per transistor is rising at sub-three-nanometer nodes, the incentive to keep coordinating around the old target... I don't know what holds that together.
Marcus Vale: Good place to stop turning it over, I think. Genuinely don't know where it lands. Thanks for pulling me into this one.
